Deep Learning with Pytorch and Hugging Face

• You can access the slide deck that covers Pytorch Here
• You can access the slide deck that covers various concepts related to Transformers Here
• It is recommended to read the slide decks before using the following colab notebooks
• Once you get a good grip on the first four modules, you can easily walk through the documentation or other code to build an application. I will keep updating this repository.
• Recorded videos

Colab Notebooks

1. The Fuel: Tensors

   • Difficulty Level: Easy if you have prior experience using Numpy or TensorFlow
   • Understand the Pytorch architecture
   • Create Tensors of 0d,1d,2d,3d,... (a multidimensional array in numpy)
   • Understand the attributes: storage, stride, offset, device
   • Manipulate tensor dimensions
   • Operations on tensors

2. The Engine: Autograd

   • Difficulty Level: Hard, requires a good understanding of backprop algorithm. However, you can skip this and still follow the subsequent notebooks easily.
   • A few more attributes of tensor : requires_grad, grad, grad_fn, _saved_tensors, backward, retain_grad, zero_grad
   • Computation graph: Leaf node (parameters) vs non-leaf node (intermediate computation)
   • Accumulate gradient and update with context manager (torch.no_grad)
   • Implementing a neural network from scratch

3. The factory: nn.Module, Data Utils

   • Difficulty Level: Medium
   • Brief tour into the source code of nn.Module
   • Everything is a module (layer in other frameworks)
   • Stack modules by subclassing nn.Module and build any neural network
   • Managing data with dataset class and DataLoader class

4. Convolutional Neural Network Image Classification

   • Difficulty Level: Medium
   • Using torchvision for datasets
   • build CNN and move it to GPU
   • Train and test
   • Transfer learning
   • Image segmentation

---

Update

• You can use various learning rate schedulers such as ExponentialLR, CosineAnnealing and so on. You just need to call scheduler.step() after optimizer.step. Refer to the documentation here
• A slight change in instantiating pre-trained models Refer

---

5. Recurrent Neural Network Sequence classification
   • Difficulty Level: Hard for pre-processing part, Medium for model building part
   • torchdata
   • torchtext
   • Embedding for words
   • Build RNN
   • Train,test, infer

---

Please take a look at the official tutorial series if you want to perform distributed training using a multi-GPU or multi-node setup in PyTorch (requires minimal modifications to the existing code). It covers various approaches, including:

• Distributed Data-Parallel (DDP)
• Fully Sharded Data Parallel (FSDP)
• Model, Tenosr and PipeLine parallelism
  Now, let's move on to the Hugging Face library, which further simplifies these training strategies

---

6. Using pre-trained models Notebook
   • Difficulty Level: Easy
   • AutoTokenizer
   • AutoModel
7. Fine-Tuning Pre-Trained Models Notebook
   • Difficulty Level: Medium
   • datasets
   • tokenizer
   • data collator with padding
   • Trainer
8. Loading Datasets Notebook
   • Difficulty Level: Easy
   • Dataset from local data files
   • Dataset from Hub
   • Preprocessing the dataset: Slice, Select, map, filter, flatten, interleave, concatenate
   • Loading from external links
9. Build a Custom Tokenizer for translation task Notebook
   • Difficulty Level: Medium
   • Translation dataset as running example
   • Building the tokenizer by encapsulating the Normalizer, pre-tokenizer and tokenization algorithm (BPE)
   • Locally Save and Load the tokenizer
   • Using it in the Transformer module
   • Exercise: Build a Tokenizer with shared vocabulary.
10. Training Custom Seq2Seq model using Vanilla Transformer Architecture Notebook
    • Difficulty Level: Medium, if you know how to build models in PyTorch.
    • Build Vanilla Transformer architecture in Pytorch
    • Create a configuration file for a model using PretrainedConfig class
    • Wrap it by HF PreTrainedModel class
    • Use the custom tokenizer built in the previous notebook
    • Use Trainer API to train the model
11. Gradient Accumulation - Continual Pre-training Notebook
    • Difficulty Level: Easy
    • Understand the memory requirement for training and inference
    • Understand how gradient accumulation overcomes the limited memory

---

CUDA Resources

Pytorch updated Cuda Semantics page on Aug 07 2025. If you are using Multiple GPUs, you must read it before starting to write code. Don't assume!